[LLM]support QWenVL second part

[DanGuge]

PR types

PR changes

Description

• 在qwen/modeling.py中添加了QWenForQWenVLInferenceModel
  • 添加了qwen-vl对应的generate_text_with_image_features逻辑
  • 重写了to_static函数
• 用户可以通过以下步骤导出模型静态图
  1. 转换qwen-vl/qwen-vl-7b权重的key
     • qwen-vl/qwen-vl-7b中权重的key与paddlenlp模型加载需要的key存在差异，需要转换
     • 下载qwen-vl/qwen-vl-7b模型，拷贝一份qwen-vl/qwen-vl-7b-inference

  import paddle
  
  state_dict = paddle.load("/path/to/.paddlenlp/models/qwen-vl/qwen-vl-7b-inference/model_state.pdparams")
  new_state_dict = {}
  for key in state_dict.keys():
      if key.startswith("transformer"):
          new_key = key.replace("transformer", "qwen")
          new_state_dict[new_key] = state_dict[key]
      else:
          new_state_dict[key] = state_dict[key]
  paddle.save(new_state_dict, "/path/to/.paddlenlp/models/qwen-vl/qwen-vl-7b-inference/model_state.pdparams")

  2. 导出静态图

  #!/bin/bash
  
  export CUDA_VISIBLE_DEVICES=0
  export PYTHONPATH=/path/to/PaddleNLP/
  
  python export_model.py \
      --model_name_or_path "/path/to/.paddlenlp/models/qwen-vl/qwen-vl-7b-inference/" \
      --output_path ./checkpoints/encode_text/ \
      --dtype float16 \
      --inference_model \
      --model_prefix qwen \
      --model_type qwen-img2txt

[paddle-bot]

Thanks for your contribution!

[codecov]

Codecov Report

Attention: 37 lines in your changes are missing coverage. Please review.

Comparison is base (c1ccafa) 56.67% compared to head (734de07) 56.66%.
Report is 2 commits behind head on develop.

Files	Patch %	Lines
...ddlenlp/experimental/transformers/qwen/modeling.py	0.00%	37 Missing ⚠️

Additional details and impacted files

@@             Coverage Diff             @@
##           develop    #7808      +/-   ##
===========================================
- Coverage    56.67%   56.66%   -0.02%     
===========================================
  Files          588      588              
  Lines        89243    89276      +33     
===========================================
+ Hits         50580    50586       +6     
- Misses       38663    38690      +27     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[wj-Mcat]

代码非常好，留了几个 comment，麻烦有时间看看。

另外，麻烦针对于这里添加一下VL 模型的单测，模型的 forward 和 动转静的过程都是需要编写的。

[wj-Mcat]

这里没有传入 input_ids，所以没必要描述与 input_ids 之间的关联吧，你这个 error message 会给开发者一头雾水。

[DanGuge]

这里是在img2txt的模型中，在第一次进入QWenInferenceModel的forward的时候，generate中不会传入input_ids，所以需要根据inputs_embeds生成一个fake的input_ids，input_ids根据bos_token_id来生成

[wawltor]

LGTM